Effect of Mass Treatment with Azithromycin on Causes of Death in Children in Malawi: Secondary Analysis from the MORDOR Trial

INTRODUCTION

Mass drug administration (MDA) with azithromycin is widely used by trachoma control programs as part of efforts to eliminate trachoma, the leading infectious cause of blindness globally.¹ During interventions principally aimed at eliminating trachoma, evidence emerged of reductions in a number of infectious diseases following azithromycin MDA, including diarrhea,² pneumonia,³ and malaria.^4–6 A study from Ethiopia reported a large reduction in child mortality following azithromycin MDA.⁷ Expert opinion suggested that a reduction in mortality was likely but to a lesser extent than estimated in this single available trial.⁸ MORDOR was a large multicenter trial that recently tested the hypothesis that biannual azithromycin MDA to children aged 1–59 months would reduce childhood mortality, reporting a 14% overall reduction in mortality in those assigned to azithromycin versus placebo.⁹ There is limited literature related to the mechanism through which azithromycin MDA may reduce child mortality. As a broad-spectrum antibiotic with weak antimalarial activity that has shown reductions in morbidity from pneumonia, diarrhea, and malaria, an effect on mortality from any of these major causes is feasible. In addition, azithromycin may reduce opportunistic infections in HIV-positive individuals, including Mycobacterium avium complex, and investigation is underway regarding its effect on HIV-associated chronic lung disease.^10,11

Ideally, the underlying cause of death would be determined from medical certificates of cause of death completed by physicians with access to adequate investigative resources. However, at the MORDOR-Malawi study site, as in low-income settings in much of sub-Saharan Africa, there was limited access to physicians who would be able to complete death certificates. In such contexts, verbal autopsy (VA)—a structured interview with relatives of the deceased to ascertain cause of death—is increasingly used to infer likely causes of death. Development of standard electronic questionnaires on mobile devices and automated analysis programs have improved the feasibility of VA and comparability of results, compared with the previous requirement for physician review of the interviews.¹² Prediction of cause-specific mortality fractions in a population using VA may be approximately 60–80% accurate, although lower for individual-level diagnosis.^13,14 In settings where more accurate determination of cause of death is not possible, VA may be used to produce probable cause of death data with the caveat of it being an imperfect tool.¹⁵ The two most commonly used VA analysis programs are InterVA¹⁶ and SmartVA.¹⁷ InterVA uses a probabilistic model based on the relationship between indicators and causes, as captured through expert panel discussions.¹⁸ SmartVA, or the Tariff method, is based on the symptom–cause information from the Population Health Metrics Research Consortium (PHMRC) study that included more than 12,000 VA interviews performed on deaths with a gold standard diagnosis derived from stringent diagnostic criteria.^19,20

The MORDOR trial was designed to assess overall mortality over three country sites and cause of death has not previously been compared between treatment arms for this study. Although azithromycin MDA has been shown to reduce morbidity due to the leading causes of child mortality outlined previously, there is no evidence to date for a cause-specific mortality effect of azithromycin MDA. As interest and evidence increases regarding the mortality benefit of azithromycin, an understanding of the mechanism for that benefit remains unclear. This study used VA methods to assess causes of death in children enrolled in the MORDOR trial in Malawi.

METHODS

RESULTS

In the 334 randomization units in the MORDOR trial in Malawi, there were 133,772 person-years of follow-up for children aged 1–59 months at the prior census. Person-years of follow-up and mortality rates were similar between males and females at each round, shown in Table 1. Over the course of the study, there were 1,184 deaths and VAs were completed for 1,131 of these. Median delay from death to VA was 5 months (range 1–26 months), with 87% of VAs completed within 12 months of death.

Mortality was 9% lower in azithromycin-treated communities compared with those allocated to placebo (rate ratio 0.91 [95% CI: 0.79–1.05]; P = 0.20). The top 10 causes of death using the InterVA algorithm are shown in Table 2, and using the SmartVA algorithm in Table 3. Malaria was the leading cause of death overall, followed by HIV/AIDS, using both algorithms without redistribution of undetermined cases, and diarrhea, pneumonia, and either acute abdomen (InterVA) or other digestive diseases (SmartVA) made up the top five causes. The top 10 causes including redistribution of undetermined cases for SmartVA are also shown in Table 3. Pneumonia, malaria, diarrhea, and HIV/AIDS accounted for 83% and 78% of the inferred causes of death using InterVA and SmartVA with redistribution, respectively.

The ITT analyses of cause-specific mortality rates are shown using InterVA in Table 4 and SmartVA in Table 5. The analysis using InterVA showed 30% lower HIV/AIDS mortality in azithromycin-treated communities (rate ratio 0.70 [95% CI: 0.50–0.97]; P = 0.03), but this trend was not evident using SmartVA (rate ratio 0.99 [95% CI: 0.71–1.38]; P = 0.95) without redistribution and rate ratio 0.92 with redistribution. The effect estimate for pneumonia mortality was lower in azithromycin communities using both InterVA (rate ratio 0.82 [95% CI: 0.60–1.12]; P = 0.22) and SmartVA (rate ratio 0.58 [95% CI: 0.33–1.00]; P = 0.05) without redistribution and rate ratio 0.87 with redistribution.

Inferred malaria mortality was similar in azithromycin communities compared with placebo using InterVA (rate ratio 0.95 [95% CI: 0.78–1.16]; P = 0.64) and SmartVA (rate ratio 0.93 [95% CI: 0.76–1.14]; P = 0.49) without redistribution and rate ratio 0.94 with redistribution. The effect estimate for diarrhea mortality was similar in azithromycin communities compared with placebo when analyzed using InterVA (rate ratio 0.95 [95% CI: 0.61–1.49]; P = 0.84) but was lower in azithromycin communities when analyzed using SmartVA (rate ratio 0.71 [95% CI: 0.51–1.00]; P = 0.05) without redistribution and rate ratio 0.72 with redistribution. The analyses of cause-specific mortality rates PP are shown in Supplemental Table 1 using InterVA, and Supplemental Table 2 using SmartVA. Cause-specific rate ratios PP were similar to those by ITT; within 10% for all causes.

Cause-specific mortality rates for the leading causes of child mortality are shown by follow-up period in Figure 1. Pneumonia mortality rates were lower in azithromycin than placebo communities at three of the four follow-up visits when analyzed using InterVA and at all follow-up visits using SmartVA. HIV/AIDS mortality appears generally lower in azithromycin communities using InterVA and diarrhea mortality generally lower using SmartVA; no such patterns are evident for malaria mortality.

View Full Size

View Full Size

Cause-specific mortality rates by intention-to-treat for the leading causes of child mortality over the four follow-up periods of the study.

Citation: The American Journal of Tropical Medicine and Hygiene 103, 3; 10.4269/ajtmh.19-0613

• Download Figure
• Download Figure
• Download figure as PowerPoint slide

The comparison of VA response terms is shown in Table 6. In the ITT analysis, the open response term “pneumonia” was more frequently reported in VAs conducted in placebo communities than azithromycin: 8.2% versus 5.1%, respectively; P = 0.03. The open response term “diarrhea” was also more frequently reported in placebo than azithromycin communities: 21.9% versus 16.9%, respectively; P = 0.03. The VA item “frequent loose/liquid stool continuing until death” was more frequently endorsed in placebo than azithromycin communities: 37.2% versus 30.1%, respectively; P = 0.01. The open response term “malaria” was similar between placebo and azithromycin arms: 45.7% versus 49.1%, respectively; P = 0.26. “Maternal test positive for HIV,” “very severe cough,” and “severe fever” were all similar between treatment arms. The PP analyses showed similar trends to the ITT results.

Assessment of seasonality by ITT using InterVA and SmartVA did not identify clear trends for differences in cause-specific mortality between treatment arms; these analyses are included as Supplemental Data. Analysis using the InterVA algorithm suggested an effect of azithromycin on pneumonia mortality during the wet season but not the dry (wet season rate ratio 0.64 [95% CI: 0.41–1.00]; P = 0.05) and dry season rate ratio 1.02 [95% CI: 0.65–1.61]; P = 0.93) (Supplemental Table 3). Analysis using SmartVA, however, predicted approximately two-thirds fewer pneumonia deaths overall and the same seasonal trend was not evident (wet season rate ratio 0.71 [95% CI: 0.32–1.61]; P = 0.42 and dry season rate ratio 0.48 [95% CI: 0.23–1.02]; P = 0.06) (Supplemental Table 4). Analysis using InterVA suggested lower HIV/AIDS mortality in azithromycin communities during both the wet and dry seasons, whereas the SmartVA analysis suggested lower diarrhea mortality in azithromycin communities during both seasons. The PP analyses using InterVA and SmartVA show generally similar results to the ITT analysis and are included for completeness as Supplemental Tables 5 and 6, respectively.

DISCUSSION

This study investigated the effect of azithromycin MDA on each of the major causes of child mortality in Mangochi District, Malawi, during the MORDOR trial. Verbal autopsy was used to infer the likely causes of death. The relative effects of azithromycin on cause-specific mortality during the wet and dry seasons was also assessed. The effect estimates produced in this study are mostly compatible with there being zero effect but are the only estimates available to date for the effect of azithromycin MDA on cause-specific child mortality.

The effect estimates using the InterVA analysis algorithm suggest lower pneumonia and HIV/AIDS mortality in azithromycin-treated than in placebo communities. Using SmartVA, the results suggest lower pneumonia and diarrhea mortality in azithromycin communities. The analysis of VA responses indicates that the terms “pneumonia” and “diarrhea” in the open narrative and “frequent loose/liquid stool continuing until death” were less commonly reported following deaths in azithromycin-treated than placebo communities. The relatively low sensitivity of InterVA for predicting both HIV/AIDS and diarrhea reported by Murray et al.,^14,19 and shown in Supplemental Table 7, indicates the algorithm may underestimate mortality from these causes. Compared with SmartVA, InterVA did predict fewer diarrhea deaths, although more HIV/AIDS deaths in this study. The relatively higher sensitivity and lower specificity of InterVA for identifying pneumonia may lead to an overestimate of pneumonia mortality. SmartVA, on the other hand, has a particularly low sensitivity for pneumonia, which would contribute to the lower proportions of pneumonia reported in this study using the SmartVA algorithm, especially before redistribution of unknown causes.

The differences in sensitivity and specificity of the two algorithms reflect the fact that VA is a blunt diagnostic tool, less suited to individual-level diagnosis compared with population-level prediction of cause-specific mortality fractions. Nonetheless, primary analyses by the two algorithms are not necessarily contradictory, both suggesting lower pneumonia mortality in azithromycin communities and InterVA suggesting lower HIV/AIDS mortality, whereas SmartVA suggests lower diarrhea mortality. These diagnoses are particularly difficult to distinguish retrospectively in the absence of clinical investigation; diarrhea may often be the immediate cause of death when the underlying cause is HIV/AIDS. The direct analysis of VA responses would suggest lower pneumonia and diarrhea in the azithromycin group; the similar maternal HIV positivity between groups, estimating HIV exposure, does not provide evidence for whether or not the diarrhea is likely to be HIV-related. The PP analyses did not show any greater effect than the ITT analyses for pneumonia, diarrhea, or HIV/AIDS, providing no evidence that these were solely individual-level effects in those treated, as opposed to community-level effects.

An effect of azithromycin on respiratory and gastrointestinal infections, including on a background of HIV exposure or infection, leading to a reduction in the rate of pneumonia and diarrhea as immediate causes of death, is certainly plausible. Azithromycin is an effective treatment for community-acquired pneumonia, including the major etiological causes, Streptococcus pneumoniae and Haemophilus influenzae.²³ Single-dose azithromycin has been shown to reduce oropharyngeal carriage of S. pneumoniae and decrease incidence of pneumonia.^24,25

Azithromycin has efficacy against the main bacterial causes of fatal childhood diarrhea, namely, enteropathogenic and enterotoxigenic Escherichia coli, Shigella spp., Campylobacter spp., Salmonella spp., and Vibrio cholerae.²⁶ Furthermore, azithromycin has anti-inflammatory properties that, combined with a reduction in gut pathogens, may reduce chronic immunostimulation caused by environmental enteric dysfunction and consequently improve nutritional status.^27,28 Evidence from the MORDOR Niger study site indicates there may be changes to the gut microbiome following azithromycin MDA, although improvements in anthropometry measurements, as an assessment of nutritional status, have not previously been associated with azithromycin MDA.^29–31

The REALITY trial identified a 27% reduction in mortality over 24 weeks in adults and children aged 5 years and older when commencing antiretroviral therapy (ART) plus an enhanced antimicrobial prophylactic regimen of trimethoprim–sulfamethoxazole, isoniazid–pyridoxine, fluconazole, albendazole, and a 5-day course of azithromycin, compared with commencing ART plus standard prophylaxis of trimethoprim–sulfamethoxazole only.³² The multiple dimensions to the enhanced regimen make it difficult to identify the specific intervention causing the reduction in mortality, but further analysis by Post et al.³³ indicated mortality reductions from cryptococcosis and unknown causes, and not from severe bacterial infections, potentially azithromycin-responsive infections or tuberculosis.

REALITY investigated mortality in adults and older children during ART commencement, so it is not directly comparable with child mortality in the MORDOR trial; antimicrobial benefits and prophylactic effects of azithromycin against chronic disease-causing pathogens such as M. avium complex and Pneumocystis jirovecii could play a role in reducing HIV/AIDS mortality in low-resource settings where HIV is underdiagnosed.³⁴ Prevalence of HIV infection in Malawi is approximately 9.2% in adults aged 15–49 years.³⁵ HIV transmission is known to be a problem in fishing communities in Mangochi District because of practices associated with the market chain.³⁶ Among HIV-infected children aged 0–14 years nationally, approximately 61% were estimated to be receiving ART in 2018.³⁵ In the setting of significant levels of undiagnosed HIV/AIDS, an antibacterial agent could prevent or delay serious infections.

In this study, malaria was the most common cause predicted by both VA analysis algorithms but there was no evidence for an effect of azithromycin MDA on malaria mortality. Azithromycin is a weak antimalarial agent, and previous studies have indicated there may be reductions in malaria morbidity following azithromycin MDA.^4–6,37 Recent assessment from the MORDOR trial in Niger indicated reduced malaria parasitemia in azithromycin-treated communities.³⁸ However, morbidity assessments as part of the MORDOR trial in Tanzania indicated no difference in fever or anemia between azithromycin and placebo arms.³⁹ A study comparing seasonal malaria chemoprevention (SMC) to SMC plus azithromycin in Burkina Faso and Mali, which have lower year-round malaria transmission than Malawi, reported no difference in mortality between groups, although reductions were evident in morbidity due to gastrointestinal and respiratory infections and nonmalarial febrile illness with the addition of azithromycin.⁴⁰

A significant limitation of this study is that MORDOR was powered to detect a difference in overall mortality at three study sites and consequently the sample size is low for this study, first because it is limited to one of the three MORDOR sites and second because it is assessing cause-specific mortality rather than all-cause mortality. The study is underpowered to provide strong conclusions regarding cause-specific mortality rates, and the small effect estimates may represent type II error. Relatively small effect sizes for the major causes of child mortality may be clinically important and this study provides first estimates for how azithromycin may reduce child mortality for hypothesis generation and further investigation.

Verbal autopsy is recognized as the only feasible method for determining the cause of death in the absence of a clinical diagnosis in low-resource settings.⁴¹ Automated analysis of VA interviews improves the affordability and reproducibility of VA compared with physician review; however, the accuracy of individual-level diagnoses should not be assumed to be as accurate as medical certification of cause of death.¹² Using the two main VA questionnaires, there is variation between the prediction of the main causes of death, so for the purposes of this investigation of the potential mechanisms of action of azithromycin, output from both algorithms was presented. The use of medical records and medical certificates of cause of death for in-facility deaths was investigated before the study but not pursued as, paradoxically, deaths that occurred in or on the way to a health facility were among the hardest to identify and categorize as there was no requirement for HSAs to report them, and health center data were generally absent or impossible to link to the community census. Any medical records retained by the family were reviewed by the VA team, and there are specific VA questions concerning diagnoses from the health system that are used in the algorithms for predicting cause of death.

In addition, VA, despite being used increasingly to estimate cause of death patterns where more sophisticated methods are not practicable, is a relatively imprecise instrument that relies on details of the final illness recalled by family members. Accuracy of recall for VA has been estimated to decrease by approximately 0.6% per month following the death, and it is recommended that, where possible, VA should be completed within 12 months.^21,42 Verbal autopsy completion for this study took place as soon as possible following identification of the death from the census, and after the customary 1-month mourning period, however, many visits were required to trace family members for some children as households in the study area move quite frequently, especially after a child’s death. The median delay of 5 months in this study and with 87% of VAs completed within 12 months is within the normally accepted range and is not expected to have significantly affected the accuracy of predicted diagnoses.

Finally, this study used the WHO 2014 VA questionnaire, which was designed to facilitate the use of both SmartVA and InterVA for assigning cause of death.⁴³ The 2014 instrument includes all questions required for InterVA, although not the exact wording for all questions from the PHMRC questionnaire used by SmartVA. A further updated questionnaire, released in 2016 after this study started, is fully compatible with both algorithms. Most questions required as input for SmartVA are similar to those in the 2014 questionnaire, although the performance of SmartVA may be slightly reduced compared with having input from the WHO 2016 questionnaire or PHMRC questionnaire. The benefits of having analysis from these two main algorithms for a wider discussion of the implications of the data were deemed to outweigh the potential for a slight decrease in performance. There is a more recently developed statistical tool for producing cause of death data from VA interviews (InSilicoVA), which aimed to improve on the other methods by sharing uncertainty between cause of death assigned for specific individuals and the population distribution of causes of death.⁴⁴ Adding a third analysis method with associated increased complexity was not attempted for this study.

Although this study is not able to provide strong evidence on the causes of death in the MORDOR trial, the data have been presented fully to enable generation of hypotheses regarding mechanisms of effect of azithromycin on child mortality. The data suggest the mortality reduction in the MORDOR trial in Malawi may have been due to effects on pneumonia and diarrhea or HIV/AIDS mortality. Larger studies will be required to clearly define the effects of azithromycin MDA on cause-specific child mortality.

Supplemental tables